@voltro/ai 0.57.0 → 0.59.0

package/dist/{inferenceQueue-CJZ4xw3R.js → inferenceQueue-CS1VouiW.js}

@@ -1,57 +1,57 @@
-import { a as e, r as t } from "./provider-BOC5wBxx.js";
-import { Output as n, generateObject as r, generateText as i, jsonSchema as a, stepCountIs as o, tool as s } from "ai";
-import { Context as c, Effect as l, JSONSchema as u, Metric as d, MetricBoundaries as ee, Runtime as f, Schema as p, Stream as m } from "effect";
-import { and as h, eq as g, generateId as _, gte as te, id as v, integer as y, json as b, queryFor as x, registerRetention as S, retentionTtlMsFromEnv as C, table as w, text as T, timestamp as E } from "@voltro/database";
-import { createHash as D } from "node:crypto";
+import { a as e, r as t, s as n } from "./provider-BOgeCRkg.js";
+import { Output as r, experimental_transcribe as i, generateObject as a, generateText as o, jsonSchema as s, stepCountIs as c, tool as l } from "ai";
+import { Context as u, Effect as d, JSONSchema as ee, Metric as f, MetricBoundaries as te, Runtime as ne, Schema as p, Stream as m } from "effect";
+import { and as h, eq as g, generateId as _, gte as re, id as v, integer as y, json as b, queryFor as x, registerRetention as S, retentionTtlMsFromEnv as C, table as w, text as T, timestamp as E } from "@voltro/database";
+import { createHash as ie } from "node:crypto";
 //#region src/tools.ts
-var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
+var ae = (e, t) => e(t), oe = Symbol.for("@voltro/ai/Tool"), se = (e) => ({
 	...e,
-	[re]: !0
-}), ae = (e) => typeof e == "object" && !!e && e[re] === !0, oe = (e) => {
+	[oe]: !0
+}), ce = (e) => typeof e == "object" && !!e && e[oe] === !0, le = (e) => {
 	let t = e.default;
 	if (typeof t != "function") return 0;
 	let n = 0;
-	for (let r of Object.values(e)) ae(r) && r.execute === void 0 && r.handler === void 0 && (r.handler = t, n += 1);
+	for (let r of Object.values(e)) ce(r) && r.execute === void 0 && r.handler === void 0 && (r.handler = t, n += 1);
 	return n;
-}, se = (e, t) => {
-	let n = t ?? f.defaultRuntime;
-	return f.runPromise(n)(e);
-}, ce = (e, t, n) => n ? e.pipe(l.timeoutFail({
+}, ue = (e, t) => {
+	let n = t ?? ne.defaultRuntime;
+	return ne.runPromise(n)(e);
+}, de = (e, t, n) => n ? e.pipe(d.timeoutFail({
 	duration: `${n} millis`,
 	onTimeout: () => /* @__PURE__ */ Error(`tool '${t}' timed out`)
-})) : e, le = (e, t, n) => {
-	if (e.execute) return se(ce(ne(e.execute, t), e.name, e.timeoutMs), n.runtime);
+})) : e, fe = (e, t, n) => {
+	if (e.execute) return ue(de(ae(e.execute, t), e.name, e.timeoutMs), n.runtime);
 	if (e.handler) {
 		let r = e.handler, i = r(t, n);
-		return l.isEffect(i) ? se(ce(i, e.name, e.timeoutMs), n.runtime) : i;
+		return d.isEffect(i) ? ue(de(i, e.name, e.timeoutMs), n.runtime) : i;
 	}
 	return Promise.reject(/* @__PURE__ */ Error(`tool '${e.name}' has neither execute nor handler`));
-}, ue = (e) => {
-	let t = u.make(e);
+}, pe = (e) => {
+	let t = ee.make(e);
 	return t && t.type === "object" ? t : {
 		type: "object",
 		properties: {},
 		additionalProperties: !1
 	};
-}, de = (e) => {
+}, me = (e) => {
 	let t = e.inputJsonSchema;
-	return t === void 0 ? ue(e.input) : t.type === "object" ? t : {
+	return t === void 0 ? pe(e.input) : t.type === "object" ? t : {
 		type: "object",
 		properties: {},
 		additionalProperties: !0
 	};
-}, fe = (e, t = { store: void 0 }) => {
+}, he = (e, t = { store: void 0 }) => {
 	let n = {};
-	for (let [r, i] of Object.entries(e)) n[i.name || r] = s({
+	for (let [r, i] of Object.entries(e)) n[i.name || r] = l({
 		description: i.description,
-		inputSchema: a(de(i)),
+		inputSchema: s(me(i)),
 		execute: async (e) => {
-			let n = await l.runPromise(p.decodeUnknown(i.input)(e)), r = await le(i, n, t);
-			return i.output ? l.runPromise(p.decodeUnknown(i.output)(r)) : r;
+			let n = await d.runPromise(p.decodeUnknown(i.input)(e)), r = await fe(i, n, t);
+			return i.output ? d.runPromise(p.decodeUnknown(i.output)(r)) : r;
 		}
 	});
 	return n;
-}, O = {
+}, D = {
 	"claude-opus-4-8": {
 		inputPer1M: 15,
 		outputPer1M: 75
@@ -96,46 +96,65 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		inputPer1M: .1,
 		outputPer1M: .4
 	}
-}, k = {}, pe = (e) => {
-	let t = k;
-	return k = { ...e }, t;
-}, A = () => ({
-	...O,
-	...k
-}), j = (e) => {
-	let t = A();
+}, O = {}, ge = (e) => {
+	let t = O;
+	return O = { ...e }, t;
+}, k = () => ({
+	...D,
+	...O
+}), A = (e) => {
+	let t = k();
 	if (t[e] !== void 0) return t[e];
 	let n = e.indexOf("/");
 	if (n > 0) {
 		let r = e.slice(n + 1);
 		if (t[r] !== void 0) return t[r];
 	}
-}, M = (e, t) => {
-	let n = e.inputTokens ?? 0, r = e.outputTokens ?? 0;
+}, j = (e) => typeof e.unit == "string", M = (e) => j(e) ? {
+	unit: e.unit,
+	input: e.input ?? 0,
+	output: e.output ?? 0
+} : {
+	unit: "tokens",
+	input: e.inputTokens ?? 0,
+	output: e.outputTokens ?? 0
+}, N = (e, t) => {
+	let { unit: n, input: r, output: i } = M(e);
 	if (t.actualCostUsd !== void 0) {
-		let e = n + r, i = e > 0 ? n / e : 0;
+		let e = r + i, a = e > 0 ? r / e : 0;
 		return {
-			inputTokens: n,
-			outputTokens: r,
-			inputCostUsd: t.actualCostUsd * i,
-			outputCostUsd: t.actualCostUsd * (1 - i),
+			inputTokens: r,
+			outputTokens: i,
+			inputCostUsd: t.actualCostUsd * a,
+			outputCostUsd: t.actualCostUsd * (1 - a),
 			totalCostUsd: t.actualCostUsd,
-			costSource: "gateway"
+			costSource: "gateway",
+			unit: n
 		};
 	}
-	let i = t.price ?? j(t.model) ?? {
+	if (n !== "tokens" && t.price === void 0) return {
+		inputTokens: r,
+		outputTokens: i,
+		inputCostUsd: 0,
+		outputCostUsd: 0,
+		totalCostUsd: 0,
+		costSource: "unpriced",
+		unit: n
+	};
+	let a = t.price ?? A(t.model) ?? {
 		inputPer1M: 0,
 		outputPer1M: 0
-	}, a = n / 1e6 * i.inputPer1M, o = r / 1e6 * i.outputPer1M;
+	}, o = r / 1e6 * a.inputPer1M, s = i / 1e6 * a.outputPer1M;
 	return {
-		inputTokens: n,
-		outputTokens: r,
-		inputCostUsd: a,
-		outputCostUsd: o,
-		totalCostUsd: a + o,
-		costSource: "estimated"
+		inputTokens: r,
+		outputTokens: i,
+		inputCostUsd: o,
+		outputCostUsd: s,
+		totalCostUsd: o + s,
+		costSource: "estimated",
+		unit: n
 	};
-}, N = "_voltro_ai_usage", P = w(N, {
+}, P = "_voltro_ai_usage", F = w(P, {
 	id: v({ prefix: "aiu" }),
 	tenantId: T().nullable(),
 	provider: T(),
@@ -146,25 +165,26 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	outputTokens: y().default(0),
 	costMicroUsd: y().default(0),
 	costSource: T().default("estimated"),
+	unit: T().default("tokens"),
 	promptId: T().nullable(),
 	promptDigest: T().nullable(),
 	promptRevision: y().nullable(),
 	calledAt: E().default("now")
-}).index(["tenantId"]).index(["promptDigest"]), me = (e) => {
+}).index(["tenantId"]).index(["promptDigest"]), _e = (e) => {
 	let t = e?.gateway?.cost, n = typeof t == "number" ? t : typeof t == "string" ? Number(t) : NaN;
 	return Number.isFinite(n) ? n : void 0;
-}, he = (e, t) => (S({
+}, ve = (e, t) => (S({
 	source: "plugin",
-	table: N,
+	table: P,
 	timeColumn: "calledAt",
 	ttlMs: C(process.env.VOLTRO_AI_USAGE_TTL_HOURS, 8760)
-}), l.promise(async () => {
-	let n = M(t.usage, {
+}), d.promise(async () => {
+	let n = N(t.usage, {
 		model: t.model,
 		...t.price === void 0 ? {} : { price: t.price },
 		...t.actualCostUsd === void 0 ? {} : { actualCostUsd: t.actualCostUsd }
 	});
-	return await e.insert(N, {
+	return await e.insert(P, {
 		tenantId: t.tenantId ?? null,
 		provider: t.provider,
 		model: t.model,
@@ -174,45 +194,46 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		outputTokens: n.outputTokens,
 		costMicroUsd: Math.round(n.totalCostUsd * 1e6),
 		costSource: n.costSource,
+		unit: n.unit,
 		promptId: t.prompt?.promptId ?? null,
 		promptDigest: t.prompt?.digest ?? null,
 		promptRevision: t.prompt?.revision ?? null,
 		calledAt: t.calledAt ?? /* @__PURE__ */ new Date()
 	}), n;
-})), ge = (e, t = {}) => l.promise(async () => {
+})), ye = (e, t = {}) => d.promise(async () => {
 	let n = [
 		...t.tenantId === void 0 ? [] : [g("tenantId", t.tenantId)],
-		...t.since === void 0 ? [] : [te("calledAt", t.since)],
+		...t.since === void 0 ? [] : [re("calledAt", t.since)],
 		...t.promptDigest === void 0 ? [] : [g("promptDigest", t.promptDigest)]
-	], r = x(P), i = n.length === 0 ? r : r.where(n.length === 1 ? n[0] : h(...n));
+	], r = x(F), i = n.length === 0 ? r : r.where(n.length === 1 ? n[0] : h(...n));
 	return (await e.query(i.descriptor)).reduce((e, t) => e + (Number(t.costMicroUsd) || 0), 0) / 1e6;
-}), F = class extends p.TaggedError()("AiBudgetExceeded", {
+}), I = class extends p.TaggedError()("AiBudgetExceeded", {
 	tenantId: p.NullOr(p.String),
 	limitUsd: p.Number,
 	spentUsd: p.Number,
 	attemptedUsd: p.Number
-}) {}, I = "_voltro_ai_budget", L = w(I, {
+}) {}, L = "_voltro_ai_budget", R = w(L, {
 	id: v({ prefix: "aib" }),
 	bucket: T().maxLength(191),
 	tenantId: T().nullable(),
 	reservedMicroUsd: y().default(0),
 	updatedAt: E().default("now")
-}).unique("byAiBudgetBucket", ["bucket"]), _e = (e, t) => `${e ?? "global"}::${t === void 0 ? "all" : String(t.getTime())}`, R = 32, ve = (e, t) => l.gen(function* () {
+}).unique("byAiBudgetBucket", ["bucket"]), be = (e, t) => `${e ?? "global"}::${t === void 0 ? "all" : String(t.getTime())}`, z = 32, xe = (e, t) => d.gen(function* () {
 	S({
 		source: "plugin",
-		table: I,
+		table: L,
 		timeColumn: "updatedAt",
 		ttlMs: C(process.env.VOLTRO_AI_USAGE_TTL_HOURS, 8760)
 	});
-	let n = t.addUsd ?? 0, r = Math.round(n * 1e6), i = Math.round(t.limitUsd * 1e6), a = _e(t.tenantId, t.since), o = t.tenantId ?? null, s = (e) => new F({
+	let n = t.addUsd ?? 0, r = Math.round(n * 1e6), i = Math.round(t.limitUsd * 1e6), a = be(t.tenantId, t.since), o = t.tenantId ?? null, s = (e) => new I({
 		tenantId: o,
 		limitUsd: t.limitUsd,
 		spentUsd: e / 1e6,
 		attemptedUsd: n
-	}), c = yield* l.promise(async () => {
-		for (let t = 0; t < R; t++) {
+	}), c = yield* d.promise(async () => {
+		for (let t = 0; t < z; t++) {
 			t > 0 && await new Promise((e) => setTimeout(e, t));
-			let n = (await e.query(x(L).where(g("bucket", a)).descriptor))[0], s = n && Number(n.reservedMicroUsd) || 0;
+			let n = (await e.query(x(R).where(g("bucket", a)).descriptor))[0], s = n && Number(n.reservedMicroUsd) || 0;
 			if (s + r > i) return {
 				ok: !1,
 				reservedMicro: s
@@ -227,7 +248,7 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 				let t = _({
 					kind: "typeid",
 					prefix: "aib"
-				}, I), n = await e.insertIgnore(I, {
+				}, L), n = await e.insertIgnore(L, {
 					id: t,
 					bucket: a,
 					tenantId: o,
@@ -237,46 +258,46 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 				if (String(n.id) === t) return { ok: !0 };
 			}
 		}
-		throw Error(`requireAiBudget: contention on bucket '${a}' exceeded ${R} attempts`);
+		throw Error(`requireAiBudget: contention on bucket '${a}' exceeded ${z} attempts`);
 	});
-	if (!c.ok) return yield* l.fail(s(c.reservedMicro));
-}), ye = ee.exponential({
+	if (!c.ok) return yield* d.fail(s(c.reservedMicro));
+}), Se = te.exponential({
 	start: .001,
 	factor: 2,
 	count: 16
-}), z = d.counter("voltro_ai_calls_total", { description: "Total AI provider calls (generate + stream)." }), B = d.counter("voltro_ai_call_errors_total", { description: "AI provider calls that errored." }), V = d.histogram("voltro_ai_call_duration_seconds", ye, "AI provider call duration in seconds."), be = d.counter("voltro_ai_input_tokens_total", {
+}), B = f.counter("voltro_ai_calls_total", { description: "Total AI provider calls (generate + stream)." }), Ce = f.counter("voltro_ai_call_errors_total", { description: "AI provider calls that errored." }), V = f.histogram("voltro_ai_call_duration_seconds", Se, "AI provider call duration in seconds."), we = f.counter("voltro_ai_input_tokens_total", {
 	description: "Total prompt (input) tokens across AI calls.",
 	incremental: !0
-}), xe = d.counter("voltro_ai_output_tokens_total", {
+}), Te = f.counter("voltro_ai_output_tokens_total", {
 	description: "Total completion (output) tokens across AI calls.",
 	incremental: !0
-}), Se = d.counter("voltro_ai_cost_microusd_total", {
+}), Ee = f.counter("voltro_ai_cost_microusd_total", {
 	description: "Total estimated AI spend in micro-USD (USD × 1e6).",
 	incremental: !0
 }), H = (e, t) => ({
 	provider: e.name,
 	model: e.model ?? e.name,
 	operation: t
-}), U = (e, t, n) => e.pipe(l.tagMetrics("provider", t.provider), l.tagMetrics("model", t.model), l.tagMetrics("operation", t.operation), l.tagMetrics("status", n)), Ce = (e, t) => {
-	let n = t.inputTokens ?? 0, r = t.outputTokens ?? 0, i = M({
+}), U = (e, t, n) => e.pipe(d.tagMetrics("provider", t.provider), d.tagMetrics("model", t.model), d.tagMetrics("operation", t.operation), d.tagMetrics("status", n)), De = (e, t) => {
+	let n = t.inputTokens ?? 0, r = t.outputTokens ?? 0, i = N({
 		inputTokens: n,
 		outputTokens: r
 	}, { model: e.model }), a = Math.round(i.totalCostUsd * 1e6);
-	return l.all([
-		U(d.incrementBy(be, n), e, "ok"),
-		U(d.incrementBy(xe, r), e, "ok"),
-		U(d.incrementBy(Se, a), e, "ok")
+	return d.all([
+		U(f.incrementBy(we, n), e, "ok"),
+		U(f.incrementBy(Te, r), e, "ok"),
+		U(f.incrementBy(Ee, a), e, "ok")
 	], { discard: !0 });
-}, W = (e, t, n) => l.gen(function* () {
-	let r = yield* l.clockWith((e) => e.currentTimeMillis);
-	return yield* n.pipe(l.tap((n) => Ce(e, t(n))), l.tap(() => U(d.increment(z), e, "ok").pipe(l.zipRight(l.clockWith((e) => e.currentTimeMillis).pipe(l.flatMap((t) => U(d.update(V, (t - r) / 1e3), e, "ok")))))), l.tapError(() => U(d.increment(z), e, "error").pipe(l.zipRight(U(d.increment(B), e, "error")), l.zipRight(l.clockWith((e) => e.currentTimeMillis).pipe(l.flatMap((t) => U(d.update(V, (t - r) / 1e3), e, "error")))))), l.withSpan("voltro.ai.call", { attributes: {
+}, W = (e, t, n) => d.gen(function* () {
+	let r = yield* d.clockWith((e) => e.currentTimeMillis);
+	return yield* n.pipe(d.tap((n) => De(e, t(n))), d.tap(() => U(f.increment(B), e, "ok").pipe(d.zipRight(d.clockWith((e) => e.currentTimeMillis).pipe(d.flatMap((t) => U(f.update(V, (t - r) / 1e3), e, "ok")))))), d.tapError(() => U(f.increment(B), e, "error").pipe(d.zipRight(U(f.increment(Ce), e, "error")), d.zipRight(d.clockWith((e) => e.currentTimeMillis).pipe(d.flatMap((t) => U(f.update(V, (t - r) / 1e3), e, "error")))))), d.withSpan("voltro.ai.call", { attributes: {
 		"ai.provider": e.provider,
 		"ai.model": e.model,
 		"ai.operation": e.operation
 	} }));
-}), we = (e, t) => {
+}), Oe = (e, t) => {
 	let n = Date.now();
-	return t.pipe(m.tap((t) => t._tag === "done" ? U(d.increment(z), e, "ok").pipe(l.zipRight(U(d.update(V, (Date.now() - n) / 1e3), e, "ok")), l.zipRight(Ce(e, t.usage ?? {}))) : t._tag === "error" ? U(d.increment(z), e, "error").pipe(l.zipRight(U(d.increment(B), e, "error")), l.zipRight(U(d.update(V, (Date.now() - n) / 1e3), e, "error"))) : l.void), m.withSpan("voltro.ai.call", { attributes: {
+	return t.pipe(m.tap((t) => t._tag === "done" ? U(f.increment(B), e, "ok").pipe(d.zipRight(U(f.update(V, (Date.now() - n) / 1e3), e, "ok")), d.zipRight(De(e, t.usage ?? {}))) : t._tag === "error" ? U(f.increment(B), e, "error").pipe(d.zipRight(U(f.increment(Ce), e, "error")), d.zipRight(U(f.update(V, (Date.now() - n) / 1e3), e, "error"))) : d.void), m.withSpan("voltro.ai.call", { attributes: {
 		"ai.provider": e.provider,
 		"ai.model": e.model,
 		"ai.operation": e.operation
@@ -285,11 +306,11 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	reason: p.Literal("generation", "decode"),
 	message: p.String
 }) {}, K = (e, t, n) => {
-	let r = [e, ...t ?? []], i = (e) => n(r[e]).pipe(l.catchTag("AiError", (t) => t.reason === "generation" && e + 1 < r.length ? i(e + 1) : l.fail(t)));
+	let r = [e, ...t ?? []], i = (e) => n(r[e]).pipe(d.catchTag("AiError", (t) => t.reason === "generation" && e + 1 < r.length ? i(e + 1) : d.fail(t)));
 	return i(0);
-}, Te = (n) => K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateText"), (e) => e.usage, l.tryPromise({
+}, ke = (n) => K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateText"), (e) => e.usage, d.tryPromise({
 	try: async () => {
-		let r = e(t), a = await i({
+		let r = e(t), i = await o({
 			model: r,
 			prompt: n.prompt,
 			...n.system === void 0 ? {} : { system: n.system },
@@ -297,34 +318,34 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 			...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
 		});
 		return {
-			text: a.text,
+			text: i.text,
 			usage: {
-				inputTokens: a.usage.inputTokens,
-				outputTokens: a.usage.outputTokens,
-				totalTokens: a.usage.totalTokens
+				inputTokens: i.usage.inputTokens,
+				outputTokens: i.usage.outputTokens,
+				totalTokens: i.usage.totalTokens
 			},
-			providerMetadata: a.providerMetadata
+			providerMetadata: i.providerMetadata
 		};
 	},
 	catch: (e) => new G({
 		reason: "generation",
 		message: String(e?.message ?? e)
 	})
-}))), q = (e, t) => l.gen(function* () {
-	let n = u.make(e);
+}))), q = (e, t) => d.gen(function* () {
+	let n = ee.make(e);
 	if (n?.type !== "object") {
 		let e = n?.anyOf === void 0 ? `type:${JSON.stringify(n?.type ?? null)}` : "anyOf (e.g. Schema.Struct({}) or a union)";
-		return yield* l.fail(new G({
+		return yield* d.fail(new G({
 			reason: "generation",
 			message: `${t}: the output schema must render as a top-level JSON object (type:"object"), but got ${e}. Use a Schema.Struct({ ... }) as the output schema — wrap arrays/scalars in a field (e.g. Schema.Struct({ items: Schema.Array(...) })).`
 		}));
 	}
 	return n;
-}), Ee = (e, t) => q(e, t), De = (n) => K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateObject"), (e) => e.usage, l.gen(function* () {
-	let i = e(t), o = yield* l.tryPromise({
-		try: () => r({
-			model: i,
-			schema: a(n.jsonSchema),
+}), Ae = (e, t) => q(e, t), je = (n) => K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateObject"), (e) => e.usage, d.gen(function* () {
+	let r = e(t), i = yield* d.tryPromise({
+		try: () => a({
+			model: r,
+			schema: s(n.jsonSchema),
 			prompt: n.prompt,
 			...n.system === void 0 ? {} : { system: n.system },
 			...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
@@ -336,16 +357,16 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		})
 	});
 	return {
-		object: o.object,
+		object: i.object,
 		usage: {
-			inputTokens: o.usage.inputTokens,
-			outputTokens: o.usage.outputTokens,
-			totalTokens: o.usage.totalTokens
+			inputTokens: i.usage.inputTokens,
+			outputTokens: i.usage.outputTokens,
+			totalTokens: i.usage.totalTokens
 		},
-		providerMetadata: o.providerMetadata
+		providerMetadata: i.providerMetadata
 	};
-}))), Oe = (e) => l.gen(function* () {
-	let t = yield* De({
+}))), Me = (e) => d.gen(function* () {
+	let t = yield* je({
 		jsonSchema: yield* q(e.schema, "generateObject"),
 		prompt: e.prompt,
 		...e.system === void 0 ? {} : { system: e.system },
@@ -353,7 +374,7 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		...e.fallbacks === void 0 ? {} : { fallbacks: e.fallbacks },
 		...e.maxTokens === void 0 ? {} : { maxTokens: e.maxTokens },
 		...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
-	}), n = yield* p.decodeUnknown(e.schema)(t.object).pipe(l.mapError((e) => new G({
+	}), n = yield* p.decodeUnknown(e.schema)(t.object).pipe(d.mapError((e) => new G({
 		reason: "decode",
 		message: `model output failed schema decode: ${String(e)}`
 	})));
@@ -361,23 +382,23 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		...t,
 		object: n
 	};
-}), ke = (r) => l.gen(function* () {
-	let s = yield* l.runtime(), c = yield* q(r.schema, "generateObjectWithTools"), u = {
+}), Ne = (n) => d.gen(function* () {
+	let i = yield* d.runtime(), a = yield* q(n.schema, "generateObjectWithTools"), l = {
 		store: void 0,
-		...r.toolContext,
-		runtime: s
-	}, d = fe(r.tools, u);
-	return yield* K(r.provider ?? t(), r.fallbacks, (t) => W(H(t, "generateObjectWithTools"), (e) => e.usage, l.gen(function* () {
-		let s = e(t), u = yield* l.tryPromise({
-			try: () => i({
-				model: s,
-				prompt: r.prompt,
-				tools: d,
-				stopWhen: o(r.maxSteps ?? 8),
-				output: n.object({ schema: a(c) }),
-				...r.system === void 0 ? {} : { system: r.system },
-				...r.maxTokens === void 0 ? {} : { maxOutputTokens: r.maxTokens },
-				...r.providerOptions === void 0 ? {} : { providerOptions: r.providerOptions }
+		...n.toolContext,
+		runtime: i
+	}, u = he(n.tools, l);
+	return yield* K(n.provider ?? t(), n.fallbacks, (t) => W(H(t, "generateObjectWithTools"), (e) => e.usage, d.gen(function* () {
+		let i = e(t), l = yield* d.tryPromise({
+			try: () => o({
+				model: i,
+				prompt: n.prompt,
+				tools: u,
+				stopWhen: c(n.maxSteps ?? 8),
+				output: r.object({ schema: s(a) }),
+				...n.system === void 0 ? {} : { system: n.system },
+				...n.maxTokens === void 0 ? {} : { maxOutputTokens: n.maxTokens },
+				...n.providerOptions === void 0 ? {} : { providerOptions: n.providerOptions }
 			}),
 			catch: (e) => new G({
 				reason: "generation",
@@ -385,47 +406,105 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 			})
 		});
 		return {
-			object: yield* p.decodeUnknown(r.schema)(u.output).pipe(l.mapError((e) => new G({
+			object: yield* p.decodeUnknown(n.schema)(l.output).pipe(d.mapError((e) => new G({
 				reason: "decode",
 				message: `model output failed schema decode: ${String(e)}`
 			}))),
 			usage: {
-				inputTokens: u.usage.inputTokens,
-				outputTokens: u.usage.outputTokens,
-				totalTokens: u.usage.totalTokens
+				inputTokens: l.usage.inputTokens,
+				outputTokens: l.usage.outputTokens,
+				totalTokens: l.usage.totalTokens
 			},
-			providerMetadata: u.providerMetadata
+			providerMetadata: l.providerMetadata
 		};
 	})));
-}), Ae = class extends c.Tag("@voltro/AiService")() {}, je = {
-	generateText: (e) => Te(e),
-	generateObject: (e) => Oe(e)
-}, Me = (e) => `sha256:${D("sha256").update(e).digest("hex").slice(0, 16)}`, Ne = class extends p.TaggedError()("PromptRenderError", {
+}), Pe = class extends u.Tag("@voltro/AiService")() {}, Fe = {
+	generateText: (e) => ke(e),
+	generateObject: (e) => Me(e)
+}, Ie = async (e, t) => {
+	if ("bytes" in e) return {
+		audio: e.bytes,
+		mediaType: e.mediaType
+	};
+	if ("url" in e) {
+		let t = await fetch(e.url);
+		if (!t.ok) throw Error(`transcribe: GET ${e.url} answered ${t.status}.`);
+		let n = t.headers.get("content-type") ?? void 0;
+		return {
+			audio: new Uint8Array(await t.arrayBuffer()),
+			...n === void 0 ? {} : { mediaType: n }
+		};
+	}
+	if (t === void 0) throw Error("transcribe: `source.storageRef` needs a `resolveRef` to turn the reference into bytes.\n  → pass `resolveRef: (ref) => storage.read(ref)`. It is injected rather than imported so @voltro/ai does not depend on storage; the app owns tenancy and signing for that read.");
+	let { bytes: n, mediaType: r } = await t(e.storageRef);
+	return {
+		audio: n,
+		mediaType: r
+	};
+}, Le = (e) => {
+	if (!Array.isArray(e) || e.length === 0) return;
+	let t = [];
+	for (let n of e) {
+		let e = typeof n.startSecond == "number" ? n.startSecond : typeof n.start == "number" ? n.start : void 0, r = typeof n.endSecond == "number" ? n.endSecond : typeof n.end == "number" ? n.end : void 0, i = typeof n.text == "string" ? n.text : void 0;
+		if (e === void 0 || r === void 0 || i === void 0) continue;
+		let a = typeof n.speaker == "string" ? n.speaker : void 0;
+		t.push({
+			start: e,
+			end: r,
+			text: i,
+			...a === void 0 ? {} : { speaker: a }
+		});
+	}
+	return t.length > 0 ? t : void 0;
+}, Re = (e) => {
+	let r = e.provider ?? t();
+	return W(H(r, "transcribe"), () => ({}), d.tryPromise({
+		try: async () => {
+			let { audio: t, mediaType: a } = await Ie(e.source, e.resolveRef), o = await i({
+				model: n(r),
+				audio: t,
+				...a === void 0 ? {} : { mediaType: a },
+				...e.providerOptions === void 0 ? {} : { providerOptions: e.providerOptions }
+			}), s = e.timestamps === "none" ? void 0 : Le(o.segments);
+			return {
+				text: o.text,
+				...o.language === void 0 ? {} : { language: o.language },
+				...o.durationInSeconds === void 0 ? {} : { durationSec: o.durationInSeconds },
+				...s === void 0 ? {} : { segments: s },
+				...o.providerMetadata === void 0 ? {} : { providerMetadata: o.providerMetadata }
+			};
+		},
+		catch: (e) => new G({
+			reason: "generation",
+			message: String(e?.message ?? e)
+		})
+	}));
+}, ze = (e) => `sha256:${ie("sha256").update(e).digest("hex").slice(0, 16)}`, Be = class extends p.TaggedError()("PromptRenderError", {
 	promptId: p.String,
 	missing: p.Array(p.String),
 	message: p.String
-}) {}, Pe = /\{\{\s*([A-Za-z0-9_.]+)\s*\}\}/g, Fe = (e) => {
+}) {}, Ve = /\{\{\s*([A-Za-z0-9_.]+)\s*\}\}/g, He = (e) => {
 	let t = /* @__PURE__ */ new Set();
-	for (let n of e.matchAll(Pe)) {
+	for (let n of e.matchAll(Ve)) {
 		let e = n[1];
 		e !== void 0 && t.add(e);
 	}
 	return [...t];
-}, Ie = (e, t) => Me(`${e.length}:${e}${(t ?? "").length}:${t ?? ""}`), Le = (e) => {
-	let t = Ie(e.template, e.system), n = Fe(e.template);
+}, Ue = (e, t) => ze(`${e.length}:${e}${(t ?? "").length}:${t ?? ""}`), We = (e) => {
+	let t = Ue(e.template, e.system), n = He(e.template);
 	return {
 		...e,
 		digest: t,
 		variables: n,
-		render: (r = {}) => l.suspend(() => {
+		render: (r = {}) => d.suspend(() => {
 			let i = n.filter((e) => r[e] === void 0);
-			if (i.length > 0) return l.fail(new Ne({
+			if (i.length > 0) return d.fail(new Be({
 				promptId: e.id,
 				missing: i,
 				message: `prompt '${e.id}' needs ${i.map((e) => `{{${e}}}`).join(", ")}, which ${i.length === 1 ? "was" : "were"} not supplied.`
 			}));
-			let a = e.template.replace(Pe, (e, t) => String(r[t]));
-			return l.succeed({
+			let a = e.template.replace(Ve, (e, t) => String(r[t]));
+			return d.succeed({
 				promptId: e.id,
 				digest: t,
 				...e.label === void 0 ? {} : { label: e.label },
@@ -435,7 +514,7 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 			});
 		})
 	};
-}, Re = (e) => typeof e == "object" && !!e && typeof e.prompt == "string" && typeof e.digest == "string" && typeof e.promptId == "string", J = "_voltro_prompts", Y = w(J, {
+}, Ge = (e) => typeof e == "object" && !!e && typeof e.prompt == "string" && typeof e.digest == "string" && typeof e.promptId == "string", J = "_voltro_prompts", Y = w(J, {
 	id: v({ prefix: "prm" }),
 	promptId: T().maxLength(191),
 	digest: T().maxLength(191),
@@ -446,12 +525,12 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	templateChars: y().default(0),
 	firstSeenAt: E().default("now"),
 	lastUsedAt: E().default("now")
-}).unique("byPromptsVersion", ["promptId", "digest"]).index(["promptId"]), ze = 8760, Be = () => {
+}).unique("byPromptsVersion", ["promptId", "digest"]).index(["promptId"]), Ke = 8760, qe = () => {
 	S({
 		source: "framework",
 		table: J,
 		timeColumn: "lastUsedAt",
-		ttlMs: C(process.env.VOLTRO_AI_PROMPTS_TTL_HOURS, ze),
+		ttlMs: C(process.env.VOLTRO_AI_PROMPTS_TTL_HOURS, Ke),
 		envVar: "VOLTRO_AI_PROMPTS_TTL_HOURS",
 		label: "ai prompt versions"
 	});
@@ -464,10 +543,10 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	system: e.system ?? null,
 	firstSeenAt: e.firstSeenAt ?? null,
 	lastUsedAt: e.lastUsedAt ?? null
-}), Ve = (e, t) => l.gen(function* () {
-	Be();
-	let n = t.digest ?? Le(t).digest;
-	return yield* l.promise(async () => {
+}), Je = (e, t) => d.gen(function* () {
+	qe();
+	let n = t.digest ?? We(t).digest;
+	return yield* d.promise(async () => {
 		let r = (await e.query(x(Y).where(h(g("promptId", t.id), g("digest", n))).descriptor))[0];
 		if (r !== void 0) {
 			let t = /* @__PURE__ */ new Date();
@@ -493,10 +572,10 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		};
 		return X(await e.insertIgnore(J, o, { conflictColumns: ["promptId", "digest"] }) ?? o);
 	});
-}), He = (e, t) => l.promise(async () => {
+}), Ye = (e, t) => d.promise(async () => {
 	let n = (await e.query(x(Y).where(g("digest", t)).descriptor))[0];
 	return n === void 0 ? null : X(n);
-}), Ue = (e, t) => l.promise(async () => [...await e.query(x(Y).where(g("promptId", t)).orderBy("revision", "desc").descriptor)].sort((e, t) => (Number(t.revision) || 0) - (Number(e.revision) || 0)).map(X)), Z = "_voltro_ai_inferences", We = w(Z, {
+}), Xe = (e, t) => d.promise(async () => [...await e.query(x(Y).where(g("promptId", t)).orderBy("revision", "desc").descriptor)].sort((e, t) => (Number(t.revision) || 0) - (Number(e.revision) || 0)).map(X)), Z = "_voltro_ai_inferences", Ze = w(Z, {
 	id: v({ prefix: "aiinf" }),
 	tenantId: T().nullable(),
 	executionId: T(),
@@ -510,7 +589,11 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		"succeeded",
 		"failed"
 	]).default("pending"),
-	kind: T().oneOf(["text", "object"]).default("text"),
+	kind: T().oneOf([
+		"text",
+		"object",
+		"transcription"
+	]).default("text"),
 	provider: T(),
 	model: T(),
 	prompt: T(),
@@ -521,6 +604,7 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	maxTokens: y().nullable(),
 	providerOptions: b().nullable(),
 	jsonSchema: b().nullable(),
+	transcription: b().nullable(),
 	result: b().nullable(),
 	errorMessage: T().nullable(),
 	attempts: y().default(0),
@@ -530,10 +614,10 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	startedAt: E().nullable(),
 	completedAt: E().nullable(),
 	durationMs: y().nullable()
-}).index(["status", "enqueuedAt"]).index(["status", "leaseUntil"]).index(["executionId"]).index(["tenantId"]), Ge = 3e5, Ke = 4, Q = (e, t) => {
+}).index(["status", "enqueuedAt"]).index(["status", "leaseUntil"]).index(["executionId"]).index(["tenantId"]), Qe = 3e5, $e = 4, Q = (e, t) => {
 	let n = typeof e == "number" ? e : Number(e);
 	return Number.isFinite(n) ? Math.trunc(n) : t;
-}, $ = (e) => e == null ? null : String(e), qe = (e) => ({
+}, $ = (e) => e == null ? null : String(e), et = (e) => ({
 	id: String(e.id),
 	tenantId: $(e.tenantId),
 	executionId: String(e.executionId),
@@ -541,7 +625,7 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	runId: $(e.runId),
 	signalName: String(e.signalName),
 	stepName: String(e.stepName),
-	kind: e.kind === "object" ? "object" : "text",
+	kind: e.kind === "object" ? "object" : e.kind === "transcription" ? "transcription" : "text",
 	provider: String(e.provider),
 	model: String(e.model),
 	prompt: String(e.prompt ?? ""),
@@ -552,8 +636,9 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 	maxTokens: e.maxTokens === null || e.maxTokens === void 0 ? null : Q(e.maxTokens, 0),
 	providerOptions: e.providerOptions ?? null,
 	jsonSchema: e.jsonSchema ?? null,
+	transcription: e.transcription ?? null,
 	attempts: Q(e.attempts, 0)
-}), Je = (e) => {
+}), tt = (e) => {
 	let t = e.attempts + 1;
 	return t >= e.maxAttempts ? {
 		kind: "give-up",
@@ -563,6 +648,6 @@ var ne = (e, t) => e(t), re = Symbol.for("@voltro/ai/Tool"), ie = (e) => ({
 		attempts: t,
 		delayMs: e.retryAfterMs !== void 0 && e.retryAfterMs > 0 ? e.retryAfterMs : Math.min(e.baseDelayMs * 2 ** (t - 1), e.maxDelayMs)
 	};
-}, Ye = (e) => `voltro.ai.infer:${e}`, Xe = (e, t) => `aiinf_${D("sha256").update(`${e}\u0000${t}`).digest("hex").slice(0, 40)}`;
+}, nt = (e) => `voltro.ai.infer:${e}`, rt = (e, t) => `aiinf_${ie("sha256").update(`${e}\u0000${t}`).digest("hex").slice(0, 40)}`;
 //#endregion
-export { W as A, A as B, Oe as C, Ee as D, Te as E, L as F, fe as G, he as H, ge as I, oe as J, ie as K, P as L, N as M, F as N, H as O, O as P, M as R, je as S, ke as T, ve as U, j as V, pe as W, Ue as _, Je as a, G as b, qe as c, Ne as d, Le as f, Ie as g, He as h, We as i, I as j, we as k, ze as l, Me as m, Ke as n, Ye as o, Re as p, ae as q, Ge as r, Xe as s, Z as t, J as u, Y as v, De as w, Ae as x, Ve as y, me as z };
+export { Oe as A, _e as B, Fe as C, ke as D, Ne as E, D as F, ve as G, k as H, R as I, he as J, xe as K, ye as L, L as M, P as N, Ae as O, I as P, F as R, Pe as S, je as T, M as U, j as V, A as W, ce as X, se as Y, le as Z, Xe as _, tt as a, Re as b, et as c, Be as d, We as f, Ue as g, Ye as h, Ze as i, W as j, H as k, Ke as l, ze as m, $e as n, nt as o, Ge as p, ge as q, Qe as r, rt as s, Z as t, J as u, Y as v, Me as w, G as x, Je as y, N as z };